> ## Documentation Index
> Fetch the complete documentation index at: https://docs.aixy-gateway.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Decision models

> Classify input, select an option, or score an ordered rubric through Aixy.

Decision models return typed evaluations of supplied evidence. Use them to classify a support
request, select a workflow, estimate whether a condition holds, or score content against a rubric.
An application can ask several questions in one call and decide how to act on the answers.

Aixy accepts the Decisions and System One formats with your project API key. The **API Reference**
tab contains the request fields, limits, examples and response schemas. Decision support is a
capability of the selected provider and model; a conversational model with structured JSON output
does not automatically become a decision model.

## Choose a question

| Evaluation | Decisions | System One | Result |
| - | - | - | - |
| Yes or no | Predicate | Noul | Probability that the condition is true |
| Select an option | Choice | Choice | Selected value, with probability and confidence evidence when supplied |
| Evaluate an ordered rubric | Score | Score | Numeric score, with per-level probabilities when supplied |

Decisions accepts a text input or user messages containing text and images, and returns answers in
question order. Its choice values can be strings or booleans. System One accepts text or structured
state, uses named questions, and returns answers under the original question keys. Its choice
options are named criteria. Its score rubric can be an ordered list, or a provider-native map of
numeric scores to criteria.

Matching native formats preserve the provider contract. Aixy translates compatible requests
between the two formats, including question identity and typed choice values. Images require native
Decisions support. Provider-specific extensions that cannot be represented in the other format
return an explicit compatibility error. Numeric-key score rubrics require native System One
support and retain their original scale. Providers can impose smaller limits than Aixy's generic
contract. A refusal remains a refusal; it is never converted into a probability or score.

## Configure a connection

Use a provider-qualified model ID and a configured organization or project connection. OpenAI
Decisions is enabled for the reviewed `gpt-6-luna` model. Other OpenAI models do not inherit that
capability.

Custom HTTPS decision providers require a saved connection configured for their native decision
format. Existing custom connections retain their conversation defaults until a decision format
is configured. Credentials and the versioned base URL belong to the saved connection; inference
requests use only the Aixy key and model ID.

Project access, model blocks, guardrails, monthly request admission, request limits and concurrency
controls still apply. Decision requests are non-streaming and do not accept output-token caps.
Hard budgets reject public decision calls because there is no provider-enforced completion bound
for a safe reservation. Token-per-minute policies also require a supported bounded estimate.

## Observe and interpret the result

Decision input and output are processed transiently and excluded from Activity content capture.
Activity retains operational metadata, provider usage when supplied, latency, status and cost
attribution. Missing provider usage or pricing remains unknown.

Probabilities and confidence describe the model's evaluation. They do not guarantee accuracy or
replace evaluation against representative examples. Choose application thresholds and handling for
ambiguous answers and refusals according to your workflow.

Public decision calls follow normal request accounting. [Task-aware routing](/guides/routing) uses
the same Rust decision execution capability internally: its classification remains part of the
generation request, with separate `routing/classify` usage evidence and no additional gateway
request unit. Routing's managed Jev option remains a classifier configured by Aixy.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.